The factor-of-IID threshold for the free Ising model on regular trees

← All digests

Family 236 · Probability and statistical mechanics
Importance 4/5 · Provisional editorial rating
Sources consulted 2026-10-11

1. The question

Imagine giving every vertex of an infinite tree its own independent random number. Can the vertices use one common symmetry-respecting rule to produce correlated spins with exactly the free Ising distribution?

There is an easy way to sample that distribution if one is allowed to choose a root: toss a fair coin there, then transmit its sign down the branches with independent errors. But the infinite regular tree has no distinguished vertex. A root used by the sampling rule breaks a symmetry that the final distribution itself possesses.

Family 236 claims the exact answer to this root-free generation problem. For every degree $d\geq3$, the free ferromagnetic Ising state is a factor of independent randomness precisely up to the reconstruction threshold, including equality. The strict impossibility beyond that threshold was already known; the claimed advance is the construction throughout the remaining range.

2. Background and essential definitions

The free Ising state

Let $T_d$ be the infinite tree in which every vertex has $d$ neighbors. A tree is connected and has no cycles, so two vertices have a unique connecting path. Write

$$b=d-1,\qquad \theta=\tanh\beta,$$

where $\beta\geq0$ is inverse temperature. The parameter $b$ counts the forward branches after a first edge has been traversed.

The free zero-field Ising law can be described by broadcasting. Choose a fair spin $\sigma_o\in\{-1,1\}$ at a temporary root $o$. Along each outward edge, independently, let a child copy its parent with probability $(1+\theta)/2$, and take the opposite sign otherwise.

Despite this rooted description, the resulting law does not depend on the choice of root. On any finite connected subtree $D$, its probability is

$$\mathbb P(\sigma|_D=\eta) =\frac12\prod_{\{u,v\}\in E(D)} \frac{1+\theta\eta_u\eta_v}{2}.$$

Here $E(D)$ is the edge set and $\eta$ is a specified spin assignment. The formula is unchanged by relabeling the tree. Spins at distance $r$ have correlation $\theta^r$.

“Free” identifies a particular infinite-volume Gibbs state, corresponding to no imposed boundary spins in finite-tree approximations. It should not be confused with the plus or minus states obtained by fixing the outer boundary to one sign.

A factor of IID

Give each vertex an independent uniform label $U_v\in[0,1]$. A factor of IID is a measurable rule $\Phi$ producing a spin configuration from the whole label collection, with the same rule used everywhere and with no distinguished root or end.

Precisely, an automorphism $g$ is a bijection of the vertices preserving adjacency. Equivariance means

$$\Phi(gU)=g\Phi(U).$$

Relabeling the input tree and then applying the rule must agree with applying the rule first and relabeling the output. The rule can inspect infinitely many labels. A factor is therefore not necessarily a finite-radius algorithm or even a finitary coding, where each output is determined by some almost-surely finite random neighborhood.

Three thresholds worth distinguishing

The uniqueness regime ends at $\theta=1/b$. Above it, different Gibbs states can coexist. Reconstruction asks a different question: do spins on very distant levels retain information about the spin at a fixed vertex? For the free Ising state on a regular tree, reconstruction is possible exactly when $b\theta^2>1$.

The factor question asks whether independent local inputs can generate the entire law equivariantly. Nonreconstruction concerns information inside a sampled configuration; it does not itself provide such a generation rule.

3. Development history and previous partial results

The reconstruction transition for the free state was established in work of Bleher, Ruiz and Zagrebnov and of Ioffe, including the critical equality on regular trees. Subsequent work by Evans, Kenyon, Peres and Schulman developed the relation between broadcasting and Ising models on trees. Pemantle and Peres later gave a nonlinear electrical-capacity criterion that handles more general critical tree geometries. Regular-tree extremality, Ioffe’s result, broadcasting framework, capacity criterion.

For factors of IID, a finite-cluster construction already handled the uniqueness range $\theta\leq1/b$, including its boundary. At the other end, Sly’s obstruction, recorded and proved in Lyons’s account, ruled out a factor when $\theta>1/\sqrt b$. Correlation bounds of Backhausz, Szegedy and Virág provide a quantitative way to see this obstruction. Lyons’s account, correlation bounds.

Nam, Sly and Zhang made the key advance beyond uniqueness. They constructed factors through infinite systems of stochastic differential equations driven by independent Brownian motions. Their theorem covered sufficiently large degree and $\theta\leq c/\sqrt b$, for absolute constants in the stated regime, and proposed reaching the full reconstruction boundary. Their construction and conjecture.

The remaining task was not merely to take a parameter limit. Weak limits of factor laws need not themselves be factors. One must preserve measurable dependence on the independent input, especially at the endpoint where ordinary geometric-decay estimates stop working.

The present manuscript follows the observation framework of Nam, Sly and Zhang, claiming the construction for every $d\geq3$ all the way through equality.

4. What the new manuscripts claim

Theorem 1.1 states that, for every integer $d\geq3$ and every $\beta\geq0$,

$$\mu_{d,\beta}\text{ is a factor of IID} \quad\Longleftrightarrow\quad \tanh\beta\leq\frac1{\sqrt{d-1}}.$$

Here $\mu_{d,\beta}$ is exactly the free, zero-field, ferromagnetic Ising law just defined.

In the permitted range, the factor can be chosen Borel measurable and equivariant under every tree automorphism on every label input. This is stronger as a choice of version than saying that equivariance holds almost surely for each fixed automorphism. The manuscript supplies a passage from the latter convention to the former for the spin factor.

For example, on the degree-three tree the uniqueness boundary is $\theta=1/2$, while the claimed factor construction reaches $\theta=1/\sqrt2$. Throughout the interval between them, the model can have multiple Gibbs states even though its free state still admits the claimed generation from independent labels.

The negative direction is not presented as new. A centered variance-one vertex factor satisfies a correlation bound of the form

$$|\mathbb E[Z_oZ_v]| \leq\left(1+\frac{d-2}{d}r\right)b^{-r/2}, \qquad \operatorname{dist}(o,v)=r.$$

Ising correlations are $\theta^r$. If $\theta\sqrt b>1$, their decay is too slow to obey the bound for large $r$. This explains why the proposed positive range is the largest possible one.

No finitary coding, coding-radius tail, runtime estimate or parameter-uniform construction is claimed. At $\beta=0$, the law is already independent fair signs; the substantial issue is the correlated regime.

5. Why the advance matters

Importance: 4/5. If correct, the theorem resolves a sharp generation question for one of the most fundamental dependent processes on a nonamenable graph. It identifies the factor threshold with the reconstruction threshold throughout the regular-tree family, including the critical point.

The result clarifies the relationship between several notions of disorder. Failure of Gibbs-state uniqueness does not prevent a particular free state from being generated from IID input. Conversely, sufficiently persistent long-range information creates an obstruction. The full interval between these two phenomena becomes explicit.

Its method also matters. It turns a representation involving an already-sampled dependent configuration into an actual measurable function of independent innovations. That distinction is useful well beyond this example: a process described using independent noise need not be recoverable from that noise alone.

6. The main idea

Hide spins in noisy observations

Temporarily sample the desired Ising configuration $\sigma$, together with independent standard Brownian motions $B_v$, and form

$$X_v(t)=t\sigma_v+B_v(t).$$

Brownian motion is continuous Gaussian noise whose variance at time $t$ is $t$. The signal grows linearly, so $X_v(t)/t$ eventually reveals $\sigma_v$.

Let $m_v(t)$ be the conditional mean of $\sigma_v$ given all observations up to time $t$. Subtract the accumulated predicted drift:

$$W_v(t)=X_v(t)-\int_0^t m_v(s)\,ds.$$

These processes are independent Brownian innovations. They represent the new noise left after accounting for information already observed.

This is not yet a factor construction, because $X$ was originally built using $\sigma$. The central question is whether the entire observation process $X$, and hence the spins, can be recovered from $W$.

Compare two possible observations with the same noise

The manuscript takes two observations $X,X'$ independently conditional on the full innovation array $W$. It first establishes a compatibility statement ensuring that the common noise remains Brownian when both observation histories are considered together.

It then aims to show that the two observations agree. Conditional independence gives

$$\mathbb E[(X_v(t)-X'_v(t))^2\mid W] =2\operatorname{Var}(X_v(t)\mid W).$$

If the difference vanishes, the conditional variance is zero, so $X$ is a measurable function of $W$. Recovering $\sigma_v$ from $X_v(n)/n$ then gives the desired spins. A uniform label can encode a Brownian path, so one independent label per vertex supplies the required input.

Why equality needs extra work

Below the reconstruction boundary, squared influences along paths contain the factor $\theta^{2r}$, while the number of vertices at distance $r$ grows like $b^r$. Their product is summable when $b\theta^2<1$, allowing a standard differential inequality to force the copies to agree.

At equality, these factors exactly cancel. The manuscript replaces the missing geometric decay with attenuation from side branches, spatial cancellation in differences between the two copies, and concentration of path averages. A further small-field estimate handles the singular behavior near observation time zero.

The final comparison uses an integrable uniqueness inequality to conclude that the discrepancy remains zero. This endpoint argument is the main additional difficulty; it cannot be replaced by simply setting $b\theta^2=1$ in a strictly subcritical geometric-series estimate.

7. Scope, limitations, and open questions

The theorem concerns regular trees, degree at least three, the free zero-field state and nonnegative inverse temperature. It does not assert the same criterion for arbitrary trees, external fields, other spin systems or every Gibbs state on the same tree.

A measurable factor may use information arbitrarily far away. The result therefore does not automatically give a distributed algorithm with a bounded number of communication rounds, an exact sampler inspecting finitely many labels per output, or a practical numerical procedure. Coding radii and computational efficiency are separate questions.

The construction is made for each fixed degree and temperature. No uniform control as parameters vary is stated. Its pointwise equivariance concerns symmetry of the selected map, not a bound on the amount of information that map uses.

The claimed coincidence between nonreconstruction and the factor property belongs to this particular model. It is not a general theorem that every nonreconstructible process on a tree is a factor of IID. Extending the construction to irregular trees or obtaining stronger coding properties would require additional ideas and hypotheses.

8. Sources and further reading

M1 states the threshold and prior converse in Section 1. Its observation framework and critical comparison explain why reconstructing the observations from their innovations is the central task.

Nam, Sly and Zhang’s earlier paper is the direct predecessor. Lyons’s Factors of IID on Trees gives the broader context and cautions about limits of factor processes.

Family 231 uses a related innovation idea for uniform spanning forests, with a different recovery argument. Family 229 discusses reconstruction thresholds for larger spin alphabets, where the relation between branching and information has additional features.

Source manuscripts

• M1. OpenAI, The sharp factor-of-IID threshold for the free Ising model on regular trees (September 26, 2026).

Sources consulted on 11 October 2026 at repository revision fd4aeeb. Download the manuscript citations.